Content recommendation method and device, computer device, storage medium and program product

Abstract

This application relates to a content recommendation method, apparatus, computer device, storage medium, and computer program product. The method includes: responding to a query request sent by a user account; obtaining recommendation tags corresponding to the query request from a tag system generated based on sample tags corresponding to sample content; obtaining account feature data of the user account, the account feature data being generated based on a text graph of the user account's historical content creation, the text graph being generated based on entity annotation results, relation annotation results, and text block tags corresponding to the historical content creation; inputting the recommendation tags and the account feature data into a recommendation model to obtain the click probability of the recommendation tags; and recommending material content corresponding to the recommendation tags to the user account according to the click probability of the recommendation tags. This method can significantly improve the accuracy of content recommendation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a content recommendation method, apparatus, computer device, storage medium, and computer program product. Background Technology

[0002] With the advent of the information explosion era on the internet, massive amounts of information are disseminated at an incredibly fast pace, placing higher demands on creators—to build high-quality content more quickly. To meet the need for rapid creation, various intelligent content creation platforms based on big data technology have emerged.

[0003] Currently common intelligent content creation platforms primarily rely on web crawling technology to acquire trending topics, and then use big data technology to process the user-selected trending topics and generate corresponding articles. After the articles are created, they can be distributed with a single click for dissemination across multiple channels. However, when users create content using current intelligent content creation platforms, the trending topics they see often do not align with their original creative intentions, indicating a problem with the low accuracy of the recommendation system. Summary of the Invention

[0004] Therefore, it is necessary to provide a high-precision content recommendation method, apparatus, computer equipment, computer-readable storage medium, and computer program product to address the aforementioned technical problems.

[0005] Firstly, this application provides a content recommendation method. The method includes:

[0006] In response to a query request sent by a user account, recommended tags corresponding to the query request are obtained from a tag system, which is generated based on sample tags corresponding to sample content;

[0007] The user account's account feature data is obtained. The account feature data is generated based on the text graph of the user account's historical content creation. The text graph is generated based on the entity annotation results, relation annotation results, and text block tags corresponding to the historical content creation.

[0008] The recommended tags and the account feature data are input into the recommendation model to obtain the click probability of the recommended tags;

[0009] Based on the click probability of the recommended tags, the user account is recommended content materials corresponding to the recommended tags.

[0010] In one embodiment, the recommendation model is connected to the knowledge graph model via a cross-compression unit, and the recommendation model learns alternately from the knowledge graph model via the cross-compression unit.

[0011] The method further includes:

[0012] Obtain the recommended entity corresponding to the query request, input the recommended entity into the knowledge graph model, and input the recommended entity into the cross-compression unit through the knowledge graph model;

[0013] The step of inputting the recommendation tag and the account feature data into the recommendation model to obtain the click probability of the recommendation tag includes:

[0014] The recommendation model is used to extract features from the account feature data to obtain account features corresponding to the user account.

[0015] The recommendation tag is input into the cross-compression unit through the recommendation model to obtain the tag features of the recommendation tag. The tag features are determined by the cross-compression unit based on the recommendation tag and the recommendation entity.

[0016] The recommendation model obtains the click probability of the user account for the recommended tag based on the account characteristics and the tag characteristics.

[0017] In one embodiment, the number of text block tags is multiple; the text graph includes entity recognition results and text structure of the historical creation content; the method further includes:

[0018] The entity annotation results and the relation annotation results corresponding to the historical creation content are determined from the triplet relations, wherein the triplet relations are generated by annotating the sample content using a triplet annotation model;

[0019] Based on the entity annotation results and the relationship annotation results, generate entity recognition results corresponding to the historical creation content;

[0020] From the tagging system, determine multiple text block tags corresponding to the historical creative content;

[0021] According to the position information of the text block corresponding to each text block label, the multiple text block labels are arranged to generate the text structure of the historical creation content.

[0022] In one embodiment, the text graph further includes textual information of the historical creation content, and the method further includes:

[0023] The historical creation content is segmented into words, and text information corresponding to the historical creation content is obtained based on the segmented words. The text information includes at least one of word cloud, sentiment attribute, or citation index. The word cloud is generated based on the frequency of the segmented words, the sentiment attribute is determined based on the similarity between the segmented words and sentiment words, and the citation index is determined based on the matching degree between the segmented words and the index.

[0024] In one embodiment, the triplet annotation model includes an entity annotation model and a relation annotation model; the method further includes:

[0025] The entity annotation model is used to extract entity features from the sample content, and the extracted entity features are used to generate sample entity annotation results corresponding to the sample content.

[0026] The relationship labeling model is used to extract relationship features from the sample entity labeling results, and sample relationship labeling results corresponding to the sample entity labeling results are generated based on the extracted relationship features.

[0027] Based on the sample entity annotation results and the sample relationship annotation results, the triplet relationship is constructed.

[0028] In one embodiment, the method further includes:

[0029] Determine the sample topic and sample text block corresponding to the sample content;

[0030] Obtain the sample topic tags corresponding to the sample topic and the sample text block tags corresponding to the sample text block;

[0031] Determine the hierarchical relationship between the sample topic tags and the sample text block tags, and construct the tag system according to the hierarchical relationship.

[0032] In one embodiment, determining the hierarchical relationship between the sample topic tags and the sample text block tags includes:

[0033] The topic entity annotation results corresponding to the sample topic and the text block entity annotation results corresponding to the sample text block are determined from the triple relationship;

[0034] Based on the topic entity annotation results and the text block entity annotation results, the entity relationship annotation results between the topic entity annotation results and the text block entity annotation results are retrieved from the triple relationship;

[0035] Based on the topic entity annotation results, the text block entity annotation results, and the entity relationship annotation results, the hierarchical relationship between the sample topic tags and the sample text block tags is determined.

[0036] In one embodiment, the method further includes:

[0037] Obtain historical behavior data corresponding to the user account;

[0038] Based on the historical behavior data and the text graph, the account feature data corresponding to the user account is determined.

[0039] In one embodiment, the method further includes:

[0040] Based on the material content selected by the user account, generate creative content corresponding to the material content;

[0041] Determine the content tags corresponding to the created content;

[0042] The target recommendation account corresponding to the created content is determined based on the content tags, and the created content is sent to the target recommendation account.

[0043] In one embodiment, the method further includes:

[0044] Based on the content text graph corresponding to the created content, feature update data corresponding to the user account is generated;

[0045] The model parameters of the recommendation model are updated using the feature update data.

[0046] Secondly, this application also provides a content recommendation device. The device includes:

[0047] The tag determination module is used to respond to a query request sent by a user account and obtain recommended tags corresponding to the query request from the tag system, which is generated based on the sample tags corresponding to the sample content;

[0048] The feature acquisition module is used to acquire account feature data of the user account. The account feature data is generated based on the text graph of the user account's historical creation content. The text graph is generated based on the entity annotation results, relation annotation results, and text block labels corresponding to the historical creation content.

[0049] The probability determination module is used to input the recommendation tag and the account feature data into the recommendation model to obtain the click probability of the recommendation tag;

[0050] The content recommendation module is used to recommend material content corresponding to the recommendation tags to the user account based on the click probability of the recommendation tags.

[0051] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the content recommendation method described in any of the embodiments of the first aspect above.

[0052] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the content recommendation method described in any of the embodiments of the first aspect above.

[0053] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the content recommendation method described in any of the embodiments of the first aspect above.

[0054] The aforementioned content recommendation methods, apparatus, computer devices, storage media, and computer program products, in response to query requests sent by user accounts, obtain recommendation tags corresponding to the query requests from a tag system. This tag system is generated based on sample tags corresponding to sample content, improving the efficiency and accuracy of tag acquisition. By acquiring user account feature data, which is generated based on the text graph of the user account's historical content creation (based on entity annotations, relation annotations, and text block tags), the system can learn the user account's creative behavior based on the text graph of historical content creation, improving the accuracy of the account feature data. By inputting recommendation tags and account feature data into a recommendation model to obtain the click probabilities of the recommendation tags, and recommending content corresponding to the recommendation tags to the user account according to these click probabilities, the recommendation model can predict the user account's click behavior on recommendation tags based on the recommendation tags and account feature data, thereby significantly improving the accuracy of content recommendation. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a content recommendation method in one embodiment. Figure 1 ;

[0056] Figure 2a This is a flowchart illustrating the click probability generation step in one embodiment;

[0057] Figure 2b This is a schematic diagram of a recommendation model and a knowledge graph model in one embodiment;

[0058] Figure 3 This is a flowchart illustrating the text graph generation steps in one embodiment;

[0059] Figure 4a This is a flowchart illustrating the steps for constructing triplet relationships in one embodiment;

[0060] Figure 4b This is a flowchart illustrating the steps involved in constructing the entity database and relational database in one embodiment.

[0061] Figure 4c This is a schematic diagram of the model architecture of an entity annotation model in one embodiment;

[0062] Figure 5 This is a flowchart illustrating the steps involved in constructing a tag system in one embodiment;

[0063] Figure 6 This is a flowchart illustrating the steps for determining hierarchical relationships in one embodiment;

[0064] Figure 7 This is a flowchart illustrating the content creation and recommendation steps in one embodiment;

[0065] Figure 8a This is a flowchart of the content recommendation method in another embodiment;

[0066] Figure 8b This is a schematic diagram of the structure of an intelligent authoring platform in one embodiment;

[0067] Figure 9 This is a structural block diagram of a content recommendation device in one embodiment;

[0068] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0070] It should be noted that the user information (including but not limited to user account information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Users may choose to refuse any content recommended to them in this application.

[0071] In one embodiment, such as Figure 1As shown, a content recommendation method is provided. This embodiment illustrates the application of this method to a server. It is understood that this method can also be applied to terminals, and to systems including terminals and servers, and is implemented through interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart TVs, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. The server can be a standalone server or a server cluster consisting of multiple servers.

[0072] In this embodiment, the method includes the following steps:

[0073] Step S102: In response to the query request sent by the user account, retrieve the recommended tags corresponding to the query request from the tag system.

[0074] The tagging system can be generated based on the sample tags corresponding to the sample content. In one example, during the tagging system construction phase, the server can tag the sample content based on multiple dimensions such as attributes, industry, scenario, and behavior to determine the sample tags corresponding to the sample content.

[0075] Specifically, the server can respond to query requests sent by user accounts, parse the query requests using preset rules, and obtain the query content carried in the query request. Semantic segmentation processing is performed on the query content to obtain query keywords corresponding to the query content. The similarity between the query keywords and tags in the tag system is obtained. Based on the similarity, recommended tags corresponding to the query keywords are determined from the tag system.

[0076] In one example, if the server parses the query content as "Economy of Province A", the server can perform semantic segmentation on the query content to obtain the query keywords "Province A" and "Economy". From the tag system, the recommended tag corresponding to the query keyword "Province A" is "Region", and the recommended tag corresponding to the query keyword "Economy" is "Economy".

[0077] Step S104: Obtain the user account's account characteristic data.

[0078] The account feature data can be generated based on a text graph of a user's historical content creation. In one example, the historical content creation could be downloaded from the internet by the server, or it could be retrieved from a local database by the server.

[0079] Text graphs can be used to represent a knowledge graph generated based on entity annotations, relation annotations, and text block labels corresponding to historical content. In one example, text block labels can be determined based on at least one of the following text block contents from historical content: chapter text block content corresponding to chapters, paragraph text block content corresponding to paragraphs, or clause text block content corresponding to clauses.

[0080] The types of entity annotation results can include, but are not limited to, location entity type, company entity type, object entity type, action entity type, and any one or more of these types.

[0081] Specifically, the server can pre-collect data from user accounts to obtain their historical content creation. This historical content is then annotated with triples to determine the corresponding entity and relation annotations. Semantic segmentation is performed on the historical content to identify text blocks. Text block tags corresponding to these text blocks are retrieved from the tagging system. Following knowledge graph construction logic, the entity annotations, relation annotations, and text block tags are processed to construct a text graph of the historical content from both triple relation and tag dimensions.

[0082] Feature extraction is performed on the text graph. Based on the feature data in the text graph, account feature data for user accounts is constructed and stored in a database. A mapping between user accounts and account feature data is established using account identifiers (such as account name, identity document, etc.). Subsequently, the server can retrieve the corresponding account feature data from the database according to the account identifier after responding to a query request sent by a user account.

[0083] Step S106: Input the recommendation tags and account feature data into the recommendation model to obtain the click probability of the recommendation tags.

[0084] The recommendation model can include, but is not limited to, any one or more of the following neural network models: DKN (Deep Knowledge-aware Network), CF (Collaborative Filtering), and the MKR model for handling recommendation tasks (Multi-Task Feature Learning for Knowledge Graph Enhanced Recommendation).

[0085] Specifically, a pre-trained recommendation model can be pre-deployed on the server. Recommendation tags and account feature data are input to the pre-trained model. The model determines the tag features of the recommendation tags and the account features of the account features. The model then processes the tag and account features to determine their similarity. Based on this similarity, the model calculates the probability of a user account clicking on a recommended tag.

[0086] Step S108: Based on the click probability of the recommended tags, recommend material content corresponding to the recommended tags to the user's account.

[0087] Specifically, the server can store the mapping between content and recommended tags. After obtaining the click probability of a user account for a recommended tag, the server can compare the click probability with a preset threshold. If the click probability of a recommended tag is greater than the preset threshold, the server retrieves the content corresponding to the recommended tag from the database according to the mapping between content and recommended tags, and pushes the content corresponding to the recommended tag to the user account.

[0088] In one example, if multiple recommended tags have a click probability greater than a preset threshold, the server can sort these tags according to their click probabilities and determine their recommendation priority based on this sorting. Content corresponding to the recommended tags is then recommended to the user's account according to this priority. For example, content corresponding to higher-priority recommended tags might be displayed at the top of the recommendation screen, or displayed in bold within the recommendation screen.

[0089] In another example, when there are multiple pieces of content corresponding to a recommendation tag, the multiple pieces of content can be sorted according to the weight of the recommendation tag in the content, the recommendation priority of each piece of content can be determined, and the content can be recommended to the user account according to the recommendation priority.

[0090] The aforementioned content recommendation method, in response to query requests sent by user accounts, retrieves recommended tags corresponding to the query request from a tag system. This tag system is generated based on sample tags corresponding to sample content, improving the efficiency and accuracy of tag retrieval. Furthermore, it acquires user account feature data, generated from the text graph of the user account's historical content creation. This text graph is generated based on entity annotations, relation annotations, and text block tags corresponding to historical content creation. This allows the method to learn the user account's creative behavior based on the text graph of historical content creation, improving the accuracy of the account feature data. By inputting recommended tags and account feature data into the recommendation model, the click probability of the recommended tags is obtained. Based on the click probability of the recommended tags, content corresponding to the recommended tags is recommended to the user account. This allows the recommendation model to predict the user account's click behavior on recommended tags based on the recommended tags and account feature data, thereby significantly improving the accuracy of content recommendation.

[0091] In one embodiment, the recommendation model can be connected to the knowledge graph model via a cross-compression unit, and the recommendation model can learn alternately from the knowledge graph model via the cross-compression unit.

[0092] like Figure 2a As shown, in step S106, the recommendation tags and account feature data are input into the recommendation model to obtain the click probability of the recommendation tags, including:

[0093] Step S202: Obtain the recommended entity corresponding to the query request, input the recommended entity into the knowledge graph model, and input the recommended entity into the cross-compression unit through the knowledge graph model.

[0094] Specifically, the server can parse the query request to obtain the query content carried in the query request. Semantic analysis is performed on the query content to determine the recommended entities corresponding to the query content. Entity relationships corresponding to the recommended entities are obtained. The recommended entities and entity relationships are input data into the knowledge graph model. The recommended entities are input into the cross-compression unit through the knowledge graph model to obtain the recommended entity features. Feature extraction is performed on the entity relationships through the knowledge graph model to obtain the entity relationship features corresponding to the entity relationships. The recommended entity features and entity relationship features are processed through the knowledge graph model to obtain the associated entities corresponding to the recommended entities, output by the knowledge graph model. Among them, the associated entities are used to represent the entities at the tail nodes that have an entity relationship with the recommended entity when the recommended entity is the head node.

[0095] Step S204: Use a multilayer perceptron to extract features from the account feature data using the recommendation model to obtain the account features corresponding to the user's account.

[0096] Among them, the multilayer perceptron (MLP) is a feedforward artificial neural network model that can be used to map multiple input datasets to a single output dataset.

[0097] Specifically, the recommendation model may include a neural network branch for extracting account features. A multilayer perceptron may be deployed on this neural network branch. The multilayer perceptron on this branch extracts account features, extracting features from the account feature data to construct an account feature vector. From this vector, the account features corresponding to the user's account are determined.

[0098] Step S206: Input the recommended tags into the cross-compression unit through the recommendation model to obtain the tag features of the recommended tags.

[0099] The cross-compression unit can be used to exchange information between two tasks.

[0100] Specifically, since the recommendation tags processed by the recommendation model and the recommendation entities processed by the knowledge graph model may overlap—for example, the recommendation tag vector for the recommendation tag "economy" and the recommendation entity vector for the recommendation entity "economy" are actually two descriptions of the same object "economy"—the server can treat the tag feature extraction task of the recommendation model and the recommendation entity feature extraction task of the knowledge graph model as two separate but related tasks. A cross-compression unit is used to achieve cross-sharing of information between the tag feature extraction task and the recommendation entity feature extraction task.

[0101] For example, a neural network branch connected to the cross-compression unit can be set up in the recommendation model. The recommendation model uses this neural network branch to input recommendation tags into the cross-compression unit, enabling data sharing between the recommendation model and the knowledge graph model. The cross-compression unit performs feature extraction on the recommendation tags and entities, determining the tag features and entity features based on the recommendation tag vector and entity vector.

[0102] Step S208: Based on account characteristics and tag characteristics, the recommendation model obtains the click probability of a user account for a recommended tag.

[0103] Specifically, the recommendation model can also include a neural network branch for predicting click probabilities, and a multilayer perceptron can be deployed in this branch. The server can process account features and tag features using the multilayer perceptron on the neural network branch for predicting click probabilities in the recommendation model to obtain the click probability of a user account for a recommended tag.

[0104] In one example, such as Figure 2b The training method for the recommendation model is explained using the MKR model for handling recommendation tasks as the recommendation model:

[0105] The server can obtain sample content created by sample accounts and corresponding recommended sample tags. It determines the entity sample annotation results, relation sample annotation results, and text block sample tags corresponding to the sample content. Based on the entity sample annotation results, relation sample annotation results, and text block sample tags, it generates a sample graph of the sample content. The sample graph is used to construct account sample feature data for the sample accounts. The account sample feature data and recommended sample tags are used to train the initial recommendation model. A first multilayer perceptron is used to extract features from the account sample feature data to obtain account sample features. A cross-compression unit is used to extract features from the recommended sample tags to obtain recommended sample tag features. A second multilayer perceptron is used to process the account sample features and recommended sample tag features to predict the click probability of recommended sample tags. Based on the predicted click probability and the actual click probability of recommended sample tags, the loss parameters of the initial recommendation model are determined. The initial model parameters of the initial recommendation model are adjusted according to the loss parameters, and iterative training continues based on the adjusted model parameters until the recommendation model obtained through iterative training meets the iterative training stopping condition (e.g., the number of iterations reaches a preset number, or the loss parameter is less than a preset loss threshold), resulting in a pre-trained recommendation model.

[0106] Simultaneously, the MKR model used for knowledge graph feature extraction is adopted as the knowledge graph model. Recommended sample entities corresponding to the sample content created by the sample account, and entity sample relationships corresponding to the recommended sample entities, are obtained. The initial knowledge graph model is trained using the recommended sample entities and entity sample relationships. A third-level multilayer perceptron is used to extract features from the entity sample relationships, yielding entity sample relationship features. A cross-compression unit is used to extract features from the recommended sample entities, yielding recommended sample entity features. A fourth-level multilayer perceptron is used to process the entity sample relationship features and recommended sample entity features to predict associated entities corresponding to the recommended sample entities. Based on the predicted associated entities and the actual associated entities corresponding to the recommended sample entities, the loss parameters of the initial knowledge graph model are determined. The initial model parameters of the initial knowledge graph model are adjusted according to the loss parameters, and iterative training continues based on the adjusted model parameters until the knowledge graph model obtained through iterative training meets the iterative training termination condition (e.g., the number of iterations reaches a preset number, or the loss parameters are less than a preset loss threshold), resulting in a pre-trained knowledge graph model.

[0107] To further improve the accuracy of content recommendation, during the training iteration of the MKR model, it is preferable to first repeat the initial training model t times, and then train the knowledge graph model once.

[0108] In this embodiment, account feature data and recommendation tags are processed separately through different neural network branches in the recommendation model. A multilayer perceptron is used to extract features from the account feature data. The recommendation tags are input into the cross-compression unit. The cross-compression unit obtains the tag features of the recommendation tags based on the vector features of the recommendation tags and recommendation entities. Then, based on the account features and tag features, the click probability of the user account for the recommendation tags is obtained. The cross-compression unit can send the recommendation entity feature information of the knowledge graph model to the recommendation model and the recommendation tag feature information of the recommendation model to the knowledge graph model, so that both can obtain additional feature information from each other. This makes up for the lack of information sparsity of a single model and improves the accuracy of the click probability.

[0109] In one embodiment, the number of text block labels is multiple. The text graph may include entity recognition results and text structure of historically created content.

[0110] like Figure 3 As shown, a flowchart illustrating the steps for generating a text graph is provided, including:

[0111] Step S302: Determine the entity annotation results and relation annotation results corresponding to the historical creation content from the triplet relations.

[0112] Step S304: Generate entity recognition results corresponding to historical creation content based on entity annotation results and relationship annotation results.

[0113] The triplet relationship can be generated by annotating the sample content using a triplet annotation model.

[0114] Specifically, the server can pre-train a triplet annotation model to extract features from the sample content. Based on the triplet features obtained from the feature extraction, the server can perform triplet annotation on the sample content to obtain triplet annotation results corresponding to the sample content. Based on the triplet annotation results of several sample contents, triplet relations are obtained and stored in a database. Based on the similarity between the triplet annotation results and the word segmentation of historical created content, entity annotation results and relation annotation results corresponding to the historical created content are determined from the triplet relations. Entity extraction processing can be performed on the historical created content according to the entity annotation results and relation annotation results to generate entity recognition results corresponding to the historical created content.

[0115] Step S306: Identify multiple text block tags corresponding to the historical creation content from the tag system.

[0116] Step S308: Arrange multiple text block labels according to the position information of the text block corresponding to each text block label to generate the text structure of historical creation content.

[0117] Specifically, the server can perform semantic segmentation on historically created content to obtain multiple text blocks corresponding to the historically created content. Each text block is then segmented into words, and based on the similarity between the segmented words and the tags in the tagging system, a text block tag corresponding to each text block is obtained from the tagging system. These multiple text block tags are then arranged according to the position information of the text block corresponding to each tag within the historically created content, and the sorted text block tags are used to generate the text structure of the historically created content. The server can then use the entity recognition results and the text structure of the historically created content as a text graph of the historically created content.

[0118] In this embodiment, entity recognition results of historical creation content are generated based on entity annotation and relation annotation results of historical creation content. According to the position information of text blocks in historical creation content, multiple text block labels are arranged to generate the text structure of historical creation content. This can construct a text graph of historical creation content from the triple relation dimension and the label dimension, improve the data richness of the text graph, and further improve the accuracy of account feature data.

[0119] In one embodiment, the text graph may further include text information of historical creation content. The text graph generation method further includes: determining the text information corresponding to the historical creation content based on word segmentation.

[0120] The text information may include, but is not limited to, at least one of word clouds, sentiment attributes, or citation metrics.

[0121] Word clouds can be used to represent thematic information of historical creative content.

[0122] Emotional attributes can be used to characterize the emotional information of historical creative content. For example, commendatory creative content generally has positive emotions, while critical creative content generally has negative emotions.

[0123] Citation metrics can be used to represent information about metrics cited in historical creative content. Metrics can be derived from official data reports.

[0124] Specifically, when the text information can include a word cloud, the frequency of occurrence of the word segment corresponding to the historical creation content can be obtained, and the corresponding word cloud can be generated according to the frequency of occurrence of the word segment.

[0125] When text information can include sentiment attributes, the similarity between the word segmentation corresponding to the historical content and the sentiment words in the sentiment database can be obtained to determine the target sentiment word corresponding to the historical content. The sentiment attribute of the historical content is then determined according to the sentiment attribute of the target sentiment word.

[0126] When the text information can include citation metrics, the semantic matching degree between word segmentation and metrics corresponding to historical creation content can be obtained to determine the citation metrics corresponding to historical creation content.

[0127] In one example, the server can also use a similarity algorithm to process the entity annotation results of the historical content and the sentiment words in the sentiment database, and to process the relationship annotation results of the historical content and the sentiment words in the sentiment database. The sentiment words corresponding to the entity annotation results and the sentiment words corresponding to the relationship annotation results are used as the target sentiment words of the historical content.

[0128] In one example, when the citation metrics have a directory hierarchy, the server can obtain the recommended metrics at the directory level corresponding to the citation metrics, and construct a citation metric table corresponding to the historically created content based on the citation metrics and recommended metrics. For instance, if the server determines that the target level corresponding to the citation metrics is level one, then the metrics at the target level two will be used as the recommended metrics corresponding to the citation metrics.

[0129] In this embodiment, by segmenting the historical creative content, word clouds, sentiment attributes, and citation indicators corresponding to the historical creative content are generated. This allows for the acquisition of textual information about the historical creative content from the dimensions of segmentation frequency, segmentation sentiment, and segmentation indicators, thereby further improving the data richness of the text graph.

[0130] In one embodiment, the triplet relation includes an entity annotation model and a relation annotation model.

[0131] like Figure 4a As shown, a flowchart illustrating the steps for constructing a triplet relation is provided, including:

[0132] Step S402: Extract entity features from the sample content using an entity annotation model, and generate sample entity annotation results corresponding to the sample content based on the extracted entity features.

[0133] Step S404: Extract relation features from the sample entity annotation results using the relation annotation model, and generate sample relation annotation results corresponding to the sample entity annotation results based on the extracted relation features.

[0134] Step S406: Construct triplet relationships based on the sample entity annotation results and sample relationship annotation results.

[0135] The sample content may include, but is not limited to, one or more of the following: structured sample data, unstructured sample data, or semi-structured sample data. In one example, the sample content may be obtained by the server from the external Internet or from an internal database.

[0136] Specifically, the server can deploy pre-trained entity annotation models and relation annotation models. Sample content can be input into the entity annotation model, which extracts entity features from the sample content, constructs a content vector of the sample content, determines the entity features of the content vector, and generates sample entity annotation results corresponding to the sample content based on the extracted entity features.

[0137] The entity annotation results can be input into a relation annotation model. This model extracts relational features from multiple entity annotation results, constructs annotation vectors for each entity, determines the relationship features between any two vectors, and then generates corresponding relational annotations based on these features. Finally, knowledge fusion processing (such as entity alignment and knowledge completion) is performed on the entity annotation results and their corresponding relational annotations to construct triplet relationships.

[0138] In one example, the server can analyze sample content and generate an initial knowledge graph structure based on the data content and structure of the sample content. The initial knowledge graph structure is then processed using the RDFS semantic model (Resource Description Framework Schema) to obtain the final knowledge graph structure. Subsequently, the server can perform knowledge fusion processing (such as entity alignment and knowledge completion) on the sample entity annotation results and sample relationship annotation results based on the knowledge graph structure.

[0139] In one example, such as Figure 4bAs shown, in the intelligent creation platform, the platform's functions can be modeled based on manually defined rules (e.g., information acquisition function: using web crawler technology to acquire trending materials and provide users with material services; content creation function: using big data technology, through recommendation algorithms and similarity calculations, to provide users with intelligent draft generation, material recommendation, text polishing, text plagiarism checking, risk detection, etc.; content distribution function: after text creation, one-click distribution and multi-channel dissemination, etc.) to obtain a knowledge graph. Manually, according to the knowledge graph, triplet annotations are applied to the initial data to obtain initial entity annotation results and initial relation annotation results. The initial entity annotation results are used to train the initial entity annotation model to obtain a pre-trained entity annotation model. The initial entity annotation results and initial relation annotation results are used to train the initial relation annotation model to obtain a pre-trained relation annotation model.

[0140] The process involves acquiring sample content, cleaning it to obtain structured and unstructured sample data, and then extracting knowledge from both based on annotation models. For structured sample data, entity annotation models are used to retrieve entity annotations from predetermined locations within the data. For unstructured sample data, entity annotation models are used to perform vector transformations, and entity annotations are obtained based on the vector features. Entity data is then extracted based on these entity annotations. Relational annotation models are then used to process the entity annotations, yielding relational annotations for the structured sample data. Finally, knowledge completion is performed on the entity and relation data using a knowledge graph. The resulting data is then stored in the entity and relation databases, respectively.

[0141] The intelligent content creation platform can then apply knowledge from the entity database and relational database to perform writing assistance operations such as intelligent search, intelligent recommendation, and knowledge query. For example, when performing intelligent recommendation, the platform can determine the recommended entities corresponding to the query request from the entity database, such as related objects and events. It then recommends multimodal content related to these entities to the user account. This multimodal content can include article-type content, image-type content, and indicator dataset-type content (the correspondence between non-text content and entities is established through manual pre-labeling).

[0142] In one example, such as Figure 4cThe diagram illustrates an entity annotation model. This model can include an input layer, an intermediate Bi-LSTM (Bi-directional Long-Short Term Memory) layer, and a top-level CRF (Conditional Random Field) layer. The entity annotation model transforms the input layer into low-dimensional dense word vectors by embedding the word / character vectors. It then uses a deep recurrent neural network combined with CRF and Bi-LSTM layers to learn contextual features from both forward and backward directions, respectively, and performs sequence signal memorization and transmission. Finally, a CRF layer at the top of the entity annotation model constrains the results of the recurrent neural network.

[0143] In another example, the relation labeling model can be obtained by training the initial relation labeling model using a supervised learning method. The relation labeling model trained by supervised learning can extract effective relation features and outperforms other training methods (such as semi-supervised learning, weakly supervised learning, or unsupervised learning) in terms of model accuracy and model recall.

[0144] In this embodiment, by using pre-trained entity annotation models and relation annotation models to annotate the sample content, and constructing triplet relations based on the sample entity annotation results and sample relation annotation results of the sample content, the efficiency and accuracy of triplet relation construction can be improved.

[0145] In one embodiment, such as Figure 5 As shown, a flowchart illustrating the steps involved in constructing a tag system is provided, including:

[0146] Step S502: Determine the sample topic and sample text block corresponding to the sample content.

[0147] Specifically, from the full-text dimension of the sample content, the server can perform full-text word segmentation to determine the frequency of each word in the sample content. Based on the frequency of word occurrence, the server determines the topic keywords corresponding to the sample content. These topic keywords are then used as the topic of the sample content. From the text block dimension of the sample content, the server can perform vector transformation processing on the sample content. Based on the feature differences between the content vectors obtained after vector transformation, the server performs semantic segmentation processing on the sample content to obtain multiple text blocks within the sample content.

[0148] Step S504: Obtain the sample topic label corresponding to the sample topic and the sample text block label corresponding to the sample text block.

[0149] Step S506: Determine the hierarchical relationship between sample topic tags and sample text block tags, and construct a tag system according to the hierarchical relationship.

[0150] Specifically, the server can use the topic keywords corresponding to the sample topic as sample topic tags. It can perform text block segmentation on the sample text blocks to determine the frequency of each segment. The relevance of text block segments with frequencies greater than a preset frequency threshold to the sample content is obtained. Based on this relevance, sample text block tags are determined. Semantic analysis algorithms are used to determine the relevance between the sample topic and the sample text blocks, and based on this relevance, the hierarchical relationship between sample topic tags and sample text block tags is determined. A tag system is constructed according to the hierarchical relationship between sample topic tags and sample text block tags.

[0151] In one example, the server can also directly obtain sample topic tags and sample text block tags obtained by manually annotating the sample content.

[0152] In another example, when the tagging system is deployed on an intelligent content creation platform, the server can obtain the content created by the platform and, referring to the sample topic tags and sample text block tags provided in the above embodiments, annotate the content to obtain content tags corresponding to it. The content tags are then processed using deep learning algorithms to learn new text tag information, expand the definition of a vocabulary system suitable for the content, and update the tagging system.

[0153] In this embodiment, by determining the sample topic and sample text block corresponding to the sample content, and based on the correlation between the sample topic and sample text block, the hierarchical relationship between the sample topic and sample text block is determined, and a tag system is constructed according to the hierarchical relationship. The tag system is updated based on the user account's created content, which can expand the amount of data stored in the tag system, improve the accuracy of the tag system construction, and improve the adaptability of the tag system to the user account's creative behavior.

[0154] In one embodiment, such as Figure 6 As shown, step S506 involves determining the hierarchical relationship between sample topic tags and sample text block tags, and constructing a tag system according to the hierarchical relationship, including:

[0155] Step S602: Determine the topic entity annotation results corresponding to the sample topic and the text block entity annotation results corresponding to the sample text block from the triple relation.

[0156] Specifically, the server can store pre-constructed triple relations. The construction operation of the triple relations can be implemented with reference to the triple construction method provided in the above embodiments, and will not be described in detail here. The server can perform vector transformation processing on the sample topic and the sample text block respectively. Based on the similarity between the sample topic vector corresponding to the sample topic and the vector features of the sample entity annotation results in the triple relation, the topic entity annotation result corresponding to the sample topic is determined from the triple relation. Based on the similarity between the sample text block vector corresponding to the sample text block and the vector features of the sample entity annotation results in the triple relation, the text block entity annotation result corresponding to the sample text block is determined from the triple relation.

[0157] Step S604: Query the entity relationship annotation results between the topic entity annotation results and the text block entity annotation results in the triple relationship.

[0158] Step S606: Determine the hierarchical relationship between the sample topic tags and the sample text block tags based on the topic entity annotation results, text block entity annotation results, and entity relationship annotation results.

[0159] Specifically, the server can retrieve the entity relation annotation results corresponding to the topic entity annotation results and text block entity annotation results from the triple query. A pre-defined deep learning algorithm is used to perform feature extraction processing on the topic entity annotation results, text block entity annotation results, and entity relation annotation results, and the association relationship between the topic entity annotation results and text block entity annotation results is determined based on the entity relation annotation results.

[0160] The algorithm learns the logic for constructing association relationships using a pre-defined deep learning algorithm, and based on this logic, determines the hierarchical relationship construction logic between sample topic tags and sample text block tags. Following this hierarchical relationship construction logic, the hierarchical relationship between sample topic tags and sample text block tags is generated.

[0161] In one example, if the topic entity annotation result corresponding to the sample topic is "catering industry" and the text block entity annotation result corresponding to the sample text block is "fresh food cold chain", the server can learn through a preset deep learning algorithm that there is a relationship between the topic entity annotation result "catering industry" and the text block entity annotation result "fresh food cold chain". Therefore, it can be determined that there is a hierarchical relationship between the tag "catering industry" and the tag "fresh food cold chain". For example, if the tag "catering industry" is a first-level tag, the second-level tag under the first-level tag "catering industry" can be "fresh food cold chain".

[0162] In another example, after the tag system is built, the server can recommend different tags to user accounts based on the tag hierarchy. For instance, if the server parses a user account's query request and the query content is "Economy of Province A", it can perform semantic analysis on the query content, dividing it into the query keyword "Province A" and the query keyword "Economy".

[0163] The tags identified from the tag system corresponding to the query keyword "Province A" can include the primary tag "Region" and the secondary tag "Province" under the primary tag "Region". Similarly, the tags identified from the tag system corresponding to the query keyword "Economy" can include the primary tag "Economy" and secondary tags such as "Bonds", "Housing Prices", "Wages", "Consumption", "Investment", and "Import and Export" under the primary tag "Economy". Subsequently, the server can cross-recommend content to user accounts based on different levels of tags; for example, it can cross-recommend multiple content materials corresponding to both the secondary tag "Province" and the primary tag "Economy".

[0164] In another example, the server can first recommend content under the primary tag corresponding to the user's query request. Then, in response to the user's click or view of the content under the primary tag, it can recommend content under the secondary tag to further narrow the search scope and improve the accuracy and efficiency of content recommendation.

[0165] In this embodiment, the hierarchical relationship between sample topic tags and sample text block tags is determined by the topic entity annotation results, text block entity annotation results, and entity relationship annotation results. This can improve the accuracy of determining the hierarchical relationship, thereby improving the accuracy of subsequent tag system construction.

[0166] In one embodiment, the content recommendation method further includes: obtaining historical behavior data corresponding to a user account, and determining account feature data corresponding to the user account based on the historical behavior data and text graph.

[0167] Historical behavioral data may include, but is not limited to, user account browsing behavior data, click behavior data, and referral behavior data.

[0168] Browsing behavior data can be obtained by the server collecting data such as the browsing content, browsing time, or browsing frequency of a user's account.

[0169] Click behavior data can be obtained by the server from data such as the content clicked by a user's account, the time of click, or the frequency of clicks.

[0170] Reference behavior data can be obtained by the server collecting data on the selected material content when performing creative actions on a user account.

[0171] Specifically, the server can retrieve historical behavioral data corresponding to a user account based on the user account's account identifier. According to preset behavioral rules, it determines behavioral characteristic data corresponding to the user account from the historical behavioral data. This behavioral characteristic data, along with characteristic data from the text graph, is used as the user account's account characteristic data. The preset behavioral rules can be based on the duration of the historical behavioral data, such as determining browsing characteristic data from browsing behavior data based on the user account's browsing time; or, they can be based on the execution frequency of the historical behavior, such as determining click characteristic data from click behavior data based on the user account's click frequency.

[0172] In this embodiment, by acquiring historical behavior data of user accounts, and based on the historical behavior data and text graphs, account feature data corresponding to user accounts is determined, which enables the learning of user account behavior characteristics. Based on user account behavior feature data and text graphs, the accuracy of account feature data is further improved.

[0173] In one embodiment, such as Figure 7 As shown, content recommendation methods also include:

[0174] Step S702: Generate creative content corresponding to the material content selected by the user account.

[0175] Specifically, content generation logic can be pre-deployed on the server. In response to a user's selection of content, the server retrieves the selected content. Based on the tags of the selected content, the server reassembles and processes the content according to the content generation logic to generate corresponding creative content. This content generation logic can be determined based on time-series forecasting logic.

[0176] Step S704: Determine the content tags corresponding to the created content.

[0177] Step S706: Determine the target recommendation account corresponding to the created content based on the content tags, and send the created content to the target recommendation account.

[0178] Specifically, the server can directly determine the content tags corresponding to the created content from the tag system; alternatively, it can manually perform semantic analysis on the created content and determine the content tags based on the semantic analysis results; or it can perform full-text word segmentation on the created content and generate a content text graph corresponding to the created content based on the word segmentation. Based on the text graph corresponding to the created content, the content tags corresponding to the created content are determined. Behavioral data of the accounts to be recommended is obtained, and based on the correlation between the content tags and the behavioral data of the accounts to be recommended, the target recommendation account corresponding to the content tags is determined from the accounts to be recommended, and the currently generated created content of the user account is sent to the target recommendation account. The specific text graph generation operation can be implemented with reference to the text graph generation method provided in the above embodiments, and will not be described in detail here.

[0179] In this embodiment, by labeling the content currently generated by the user account, determining the content tags corresponding to the content, and sending the content to the target recommendation account based on the content tags, personalized push of content can be achieved, thereby improving the recommendation efficiency of content.

[0180] In one embodiment, in Figure 7 Building upon the provided content recommendation method, the method also includes: generating feature update data corresponding to user accounts based on the content text graph of the created content. The feature update data is then used to update the model parameters of the recommendation model.

[0181] Specifically, the server can process the created content using the text graph generation method provided in the above embodiments to generate a content text graph corresponding to the created content. Feature extraction processing is performed on the content text graph corresponding to the created content to determine the feature update data corresponding to the user account from the content text graph data. The feature update data is used to iteratively train the recommendation model, and the model parameters of the recommendation model are updated based on the model loss function of the iteratively trained model to obtain the updated recommendation model. Subsequently, the click probability is obtained using the iteratively trained recommendation model.

[0182] In one example, where feature update data may include tag update data, the feature update data can be determined based on the content tags of the created content. The server can use the tag update data to update the tag system, and then use the updated tag system to update the model parameters of the recommendation model, resulting in an updated recommendation model.

[0183] In another example, where the feature update data may include triple update data, the feature update data can be manually determined by labeling the created content with triples. The server can use the triple update data to update the triple relationships, and then use the updated triple relationships to update the model parameters of the recommendation model, resulting in an updated recommendation model.

[0184] In this embodiment, feature extraction is performed on the content text graph corresponding to the created content to generate feature update data corresponding to the user account. The feature update data is then used to update the model parameters of the recommendation model. This allows updating the pre-built triplet relationships and / or tag system based on the currently created content of the user account. The updated triplet relationships and / or tag system are then used to update the model parameters of the recommendation model, further improving the accuracy of the click probability generation of the recommendation model.

[0185] In one embodiment, such as Figure 8a As shown, a content recommendation method is provided, including:

[0186] Step S802: Process the sample content using entity annotation model and relation annotation model to construct triplet relations.

[0187] Specifically, the server can extract entity features from the sample content using an entity annotation model to obtain entity annotation results. Then, it can extract relation features from the entity annotation results using a relation annotation model to obtain relation annotation results. Based on the entity annotation results and relation annotation results, triplet relations are constructed. The specific triplet relation construction operation can be implemented using the triplet construction method provided in the above embodiments, and will not be elaborated upon here.

[0188] Step S804: Based on the triplet relationship between the sample topic and the sample text block, determine the hierarchical relationship between the sample topic tag and the sample text block tag, and construct the tag system.

[0189] Specifically, the server can perform semantic segmentation on the sample content to determine the sample topic and multiple sample text blocks within the sample content. From the triplet relations, the server determines the topic entity annotation results corresponding to the sample topic and the text block entity annotation results corresponding to the sample text blocks. Based on the topic entity annotation results and text block entity annotation results, the server retrieves the entity relation annotation results corresponding to the topic entity annotation results and text block entity annotation results from the triplet query, and determines the association relationship between the topic entity annotation results and the text block entity annotation results based on the entity relation annotation results. The server uses the TF-IDF (term frequency–inverse document frequency) algorithm to obtain the sample topic tags corresponding to the sample topic and the sample text block tags corresponding to the sample text blocks. Based on the association relationship between the topic entity annotation results and the text block entity annotation results, the server determines the hierarchical relationship between the sample topic tags and the sample text block tags. A tag system is constructed according to the hierarchical relationship. The specific tag system construction operation can be implemented with reference to the tag system construction method provided in the above embodiments, and will not be elaborated here.

[0190] Step S806: In response to the query request sent by the user account, retrieve the recommended tags corresponding to the query request from the tag system, and determine the recommended entities corresponding to the query request from the triple relationship.

[0191] Specifically, the server can respond to a query request sent by a user account, parsing it to obtain the query content corresponding to the query request. Semantic analysis is performed on the query content, and based on the results of this analysis, recommended tags corresponding to the query request are determined from the tag system, and recommended entities corresponding to the query request are determined from the triple relationships. The entity relationships obtained from the triple relationships are then retrieved according to the recommended entities. The specific operations for obtaining recommended tags and recommended entities can be implemented using the methods provided in the above embodiments, and will not be elaborated upon here.

[0192] In one example, the server can obtain the content materials corresponding to the recommendation tags and the recommendation entities, and directly push these materials to the user's account. Based on the user's browsing and clicking actions on the content materials, the server obtains the entity relationships corresponding to the recommendation tags and / or recommendation entities selected by the user. Subsequent click probability calculation operations are then performed based on the recommendation tags and / or recommendation entities selected by the user.

[0193] Step S808: Generate user account feature data based on the text graph of the user account's historical content creation and the user account's historical behavior data.

[0194] Specifically, the server can obtain the user's historical creation content, perform word segmentation on the historical creation content, and generate word clouds, sentiment attributes, and citation indicators corresponding to the historical creation content based on the word segmentation, thus determining the text information of the historical creation content. Entity annotation results and relation annotation results corresponding to the historical creation content are determined from the triple relations. Based on the entity annotation results and relation annotation results, the entity recognition results of the historical creation content are determined. Multiple text block tags corresponding to the historical creation content are determined from the tag system, and these tags are sorted according to the position information of the text blocks in the historical creation content to generate the text structure of the historical creation content. The word cloud, sentiment attributes, citation indicators, entity recognition results, and text structure of the historical creation content are used as the text graph of the historical creation content. The specific text graph generation operation can be implemented with reference to the text graph generation method provided in the above embodiments, and will not be described in detail here.

[0195] Step S810: Input the account feature data and recommendation tags into the recommendation model to obtain the click probability of the user account on the recommendation tags, and recommend material content to the user account according to the click probability of the recommendation tags.

[0196] Specifically, the server can input user account feature data and recommendation tags into the recommendation model, and input recommendation entities and entity relationships into the knowledge graph model. The recommendation model performs feature extraction on the account feature data to obtain the account features corresponding to the account feature data. The recommendation tag is input into the cross-compression unit through the recommendation model, and the recommendation entity is input into the cross-compression unit through the knowledge graph model. The cross-compression unit performs feature extraction on the recommendation tag and recommendation entity to generate tag features corresponding to the recommendation tag and recommendation entity features corresponding to the recommendation entity. Based on the similarity between the account features and tag features, the click probability of the user account for the recommendation tag is obtained. According to the click probability, multimodal content corresponding to the recommendation tag is recommended to the user account. The specific click probability generation operation can be implemented with reference to the click probability generation method provided in the above embodiments, and will not be elaborated here.

[0197] Step S812: Generate content based on the selected material, determine the content tags corresponding to the content created by the user account, and send the content to the target recommended account based on the content tags.

[0198] Specifically, after the server responds to a query request sent by a user account and recommends content corresponding to the recommended tags, the server can respond to the user account's selection of the recommended content and obtain the selected content. The content is then processed to generate corresponding creative content. Content tags corresponding to the user account's creative content are determined. Based on the content tags, the target recommendation account for the creative content is identified, and the creative content is sent to the target recommendation account.

[0199] The created content can also be processed to generate a content text graph corresponding to the content. The tag system is updated using the tag update data from the content text graph to obtain the updated tag system. In this case, the triplet relationships are updated using the triplet update data from the content text graph to obtain the updated triplet relationships. The updated tag system and the updated triplet relationships are then used to update the model parameters of the recommendation model to obtain the updated recommendation model. The specific model parameter update operations can be implemented using the model parameter update method provided in the above embodiments, and will not be elaborated here.

[0200] In one example, such as Figure 8b As shown, an intelligent creation platform 800 is provided, including: a knowledge map module 802, an AI (Artificial Intelligence) tag system module 804, an intelligent recommendation module 806, a creation insight module 808, and a creation feedback module 810.

[0201] Specifically, the intelligent creation platform can use the entity extraction model and relation extraction model in the knowledge map module 802 to process sample content and construct triple relationships. It can then use the AI ​​tag module 804, based on the triple relationships constructed by the knowledge map module 802, to determine the association between sample topics and sample paragraphs. Using the association between sample topics and sample paragraphs, it can determine the hierarchical relationship between sample topic tags and sample text block tags, thus constructing a tag system.

[0202] The intelligent recommendation module 806 can use the triplet relationships constructed based on the knowledge map module 802 to determine the recommended entities corresponding to the query requests sent by the user account, and the tag system constructed based on the AI ​​tag system module 804 to determine the recommended tags corresponding to the query requests sent by the user account. The intelligent recommendation module 806 uses the user account's account feature data, inputs the account feature data and recommended tags into the recommendation model to generate the click probabilities of the recommended tags, and recommends content to the user account according to the click probabilities.

[0203] The Creation Insight Module 808 can be used to generate creative content corresponding to the material content selected by the user account. The Creation Insight Module 808 then processes the user account's creative content to generate corresponding text graphs. The Creation Feedback Module 810, based on the creative content generated by the Creation Insight Module 808, determines the content tags for the creative content and sends it to the target recommendation account according to these tags. Simultaneously, feature update data is determined based on the creative content. Triple update data is sent to the Knowledge Map Module 802 to update the triple relationships, and tag update data is sent to the AI ​​Tag System Module 804 to update the tag system. The Intelligent Recommendation Module 806 updates the recommendation model based on the updated triple relationships from the Knowledge Map Module 802 and the updated tag system from the AI ​​Tag System Module 804, resulting in an updated recommendation model. The updated recommendation model is then used to generate the click probabilities of the recommended tags.

[0204] In this embodiment, in response to a query request sent by a user account, recommended tags corresponding to the query request are obtained from a tag system. This tag system is generated based on sample tags corresponding to sample content, improving the efficiency and accuracy of tag acquisition. Account feature data of the user account is obtained. This data is generated based on the text graph of the user account's historical content creation. The text graph is generated based on entity annotations, relation annotations, and text block tags corresponding to the historical content creation. This allows the system to learn the user account's creative behavior based on the text graph of historical content creation, improving the accuracy of the account feature data. By inputting the recommended tags and account feature data into the recommendation model, the click probability of the recommended tags is obtained. Based on the click probability of the recommended tags, content corresponding to the recommended tags is recommended to the user account. The recommendation model can predict the user account's click behavior on recommended tags based on the recommended tags and account feature data, thereby significantly improving the accuracy of content recommendation.

[0205] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the steps or stages of other steps.

[0206] Based on the same inventive concept, this application also provides a content recommendation apparatus for implementing the content recommendation method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more content recommendation apparatus embodiments provided below can be found in the limitations of the content recommendation method described above, and will not be repeated here.

[0207] In one embodiment, such as Figure 9 As shown, a content recommendation device 900 is provided, including: a tag determination module 902, a feature acquisition module 904, a probability determination module 906, and a content recommendation module 908, wherein:

[0208] The tag determination module 902 is used to respond to the query request sent by the user account and obtain the recommended tags corresponding to the query request from the tag system. The tag system is generated based on the sample tags corresponding to the sample content.

[0209] The feature acquisition module 904 is used to acquire the account feature data of the user account. The account feature data is generated based on the text graph of the user account's historical content creation. The text graph is generated based on the entity annotation results, relation annotation results, and text block labels corresponding to the historical content creation.

[0210] The probability determination module 906 inputs the recommended tags and account feature data into the recommendation model to obtain the click probability of the recommended tags.

[0211] The content recommendation module 908 is used to recommend material content corresponding to the recommended tags to the user's account based on the click probability of the recommended tags.

[0212] In one embodiment, the recommendation model is connected to the knowledge graph model via a cross-compression unit, and the recommendation model learns alternately from the knowledge graph model via the cross-compression unit. The content recommendation device 900 further includes an entity determination module, used to obtain the recommended entity corresponding to the query request, input the recommended entity into the knowledge graph model, and input the recommended entity into the cross-compression unit via the knowledge graph model.

[0213] The probability determination module 906 includes: an account feature extraction unit, used to extract features from account feature data using a multilayer perceptron through a recommendation model to obtain account features corresponding to user accounts; a tag feature extraction unit, used to input recommendation tags into a cross-compression unit through a recommendation model to obtain tag features of recommendation tags, the tag features being determined by the cross-compression unit based on recommendation tags and recommendation entities; and a click probability determination unit, used to obtain the click probability of a user account for a recommendation tag based on account features and tag features through a recommendation model.

[0214] In one embodiment, the number of text block labels is multiple; the text graph includes entity recognition results and text structure of historical creation content. The content recommendation device 900 further includes: a result determination module, used to determine entity annotation results and relation annotation results corresponding to historical creation content from triplet relations, wherein the triplet relations are generated by annotating sample content using a triplet annotation model; a result generation module, used to generate entity recognition results corresponding to historical creation content based on entity annotation results and relation annotation results; a text block label determination module, used to determine multiple text block labels corresponding to historical creation content from a labeling system; and a structure generation module, used to arrange multiple text block labels according to the position information of the text blocks corresponding to each text block label, generating the text structure of historical creation content.

[0215] In one embodiment, the text graph also includes text information of historical creation content. The content recommendation device 900 further includes: an information determination module, used to determine text information corresponding to historical creation content based on word segmentation, wherein the text information includes at least one of word cloud, sentiment attribute, or citation index.

[0216] In one embodiment, the triplet annotation model includes an entity annotation model and a relation annotation model. The content recommendation device 900 further includes: an entity annotation module, used to extract entity features from the sample content using the entity annotation model, and generate sample entity annotation results corresponding to the sample content based on the extracted entity features; a relation annotation module, used to extract relation features from the sample entity annotation results using the relation annotation model, and generate sample relation annotation results corresponding to the sample entity annotation results based on the extracted relation features; and a relation construction module, used to construct triplet relations based on the sample entity annotation results and the sample relation annotation results.

[0217] In one embodiment, the content recommendation device 900 further includes: a text segmentation module for determining sample topics and sample text blocks corresponding to sample content; a text tag acquisition module for acquiring sample topic tags corresponding to sample topics and sample text block tags corresponding to sample text blocks; and a system construction module for determining the hierarchical relationship between sample topic tags and sample text block tags, and constructing a tag system according to the hierarchical relationship.

[0218] In one embodiment, the system construction module includes: an entity annotation determination unit, used to determine the topic entity annotation result corresponding to the sample topic and the text block entity annotation result corresponding to the sample text block from the triple relationship; a relationship annotation unit, used to query the entity relationship annotation result between the topic entity annotation result and the text block entity annotation result from the triple relationship; and a hierarchical relationship determination unit, used to determine the hierarchical relationship between the sample topic label and the sample text block label based on the topic entity annotation result, the text block entity annotation result, and the entity relationship annotation result.

[0219] In one embodiment, the content recommendation device 900 further includes: a behavior data acquisition module for acquiring historical behavior data corresponding to a user account; and a feature data determination module for determining account feature data corresponding to a user account based on the historical behavior data and text graph.

[0220] In one embodiment, the content recommendation device 900 further includes: a content generation module, used to generate creative content corresponding to the material content selected by the user account; a content tag determination module, used to determine the content tags corresponding to the creative content; and a content sending module, used to determine the target recommendation account corresponding to the creative content based on the content tags, and send the creative content to the target recommendation account.

[0221] In one embodiment, the content recommendation device 900 further includes: a feature update module, used to generate feature update data corresponding to a user account based on the content text graph corresponding to the created content; and a model update module, used to update the model parameters of the recommendation model using the feature update data.

[0222] The modules in the aforementioned recommended device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0223] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores historical content creation, historical behavior data, triplet relationships, tagging systems, and recommendation models. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a content recommendation method.

[0224] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0225] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0226] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps in the above method embodiments.

[0227] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0228] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0229] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0230] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A content recommendation method, characterized in that, The method includes: In response to a query request sent by a user account, recommended tags corresponding to the query request are obtained from a tag system, which is generated based on sample tags corresponding to sample content; The user account's account feature data is obtained. The account feature data is generated based on the text graph of the user account's historical content creation. The text graph is generated based on the entity annotation results, relation annotation results, and text block tags corresponding to the historical content creation. The recommended tags and the account feature data are input into the recommendation model to obtain the click probability of the recommended tags; Based on the click probability of the recommended tags, recommend material content corresponding to the recommended tags to the user account; The recommendation model is connected to the knowledge graph model through a cross-compression unit, and the recommendation model learns alternately from the knowledge graph model through the cross-compression unit. The method further includes: Obtain the recommended entity corresponding to the query request, input the recommended entity into the knowledge graph model, and input the recommended entity into the cross-compression unit through the knowledge graph model; The step of inputting the recommendation tag and the account feature data into the recommendation model to obtain the click probability of the recommendation tag includes: The recommendation model uses a multilayer perceptron to extract features from the account feature data to obtain account features corresponding to the user account; The recommendation tag is input into the cross-compression unit through the recommendation model to obtain the tag features of the recommendation tag. The tag features are determined by the cross-compression unit based on the recommendation tag and the recommendation entity. The recommendation model obtains the click probability of the user account for the recommended tag based on the account characteristics and the tag characteristics.

2. The method according to claim 1, characterized in that, The number of text block tags is multiple; the text graph includes entity recognition results and text structure of the historical creation content; the method further includes: The entity annotation results and the relation annotation results corresponding to the historical creation content are determined from the triplet relations, wherein the triplet relations are generated by annotating the sample content using a triplet annotation model; Based on the entity annotation results and the relationship annotation results, generate entity recognition results corresponding to the historical creation content; From the tagging system, determine multiple text block tags corresponding to the historical creative content; According to the position information of the text block corresponding to each text block label, the multiple text block labels are arranged to generate the text structure of the historical creation content.

3. The method according to claim 2, characterized in that, The text graph also includes textual information about the historical creation content, and the method further includes: Based on the word segmentation corresponding to the historical creative content, the text information corresponding to the historical creative content is determined, and the text information includes at least one of word cloud, sentiment attribute or citation index.

4. The method according to claim 2, characterized in that, The triplet annotation model includes an entity annotation model and a relation annotation model; the method further includes: The entity annotation model is used to extract entity features from the sample content, and the extracted entity features are used to generate sample entity annotation results corresponding to the sample content. The relationship labeling model is used to extract relationship features from the sample entity labeling results, and sample relationship labeling results corresponding to the sample entity labeling results are generated based on the extracted relationship features. Based on the sample entity annotation results and the sample relationship annotation results, the triplet relationship is constructed.

5. The method according to claim 1, characterized in that, The method further includes: Determine the sample topic and sample text block corresponding to the sample content; Obtain the sample topic tags corresponding to the sample topic and the sample text block tags corresponding to the sample text block; Determine the hierarchical relationship between the sample topic tags and the sample text block tags, and construct the tag system according to the hierarchical relationship.

6. The method according to claim 5, characterized in that, Determining the hierarchical relationship between the sample topic tags and the sample text block tags includes: The topic entity annotation results corresponding to the sample topic and the text block entity annotation results corresponding to the sample text block are determined from the triple relationship; The entity relationship annotation result between the topic entity annotation result and the text block entity annotation result is obtained by querying the triple relationship; Based on the topic entity annotation results, the text block entity annotation results, and the entity relationship annotation results, the hierarchical relationship between the sample topic tag and the sample text block tag is determined.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Obtain historical behavior data corresponding to the user account; Based on the historical behavior data and the text graph, the account feature data corresponding to the user account is determined.

8. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Based on the material content selected by the user account, generate creative content corresponding to the material content; Determine the content tags corresponding to the created content; The target recommendation account corresponding to the created content is determined based on the content tags, and the created content is sent to the target recommendation account.

9. The method according to claim 8, characterized in that, The method further includes: Based on the content text graph corresponding to the created content, feature update data corresponding to the user account is generated; The model parameters of the recommendation model are updated using the feature update data.

10. A content recommendation device, characterized in that, The device includes: The tag determination module is used to respond to a query request sent by a user account and obtain recommended tags corresponding to the query request from the tag system, which is generated based on the sample tags corresponding to the sample content. The feature acquisition module is used to acquire account feature data of the user account. The account feature data is generated based on the text graph of the user account's historical creation content. The text graph is generated based on the entity annotation results, relation annotation results, and text block labels corresponding to the historical creation content. The probability determination module is used to input the recommendation tag and the account feature data into the recommendation model to obtain the click probability of the recommendation tag; The content recommendation module is used to recommend material content corresponding to the recommendation tags to the user account based on the click probability of the recommendation tags; The recommendation model is connected to the knowledge graph model through a cross-compression unit, and the recommendation model learns alternately from the knowledge graph model through the cross-compression unit. The device further includes: The entity determination module is used to obtain the recommended entity corresponding to the query request, input the recommended entity into the knowledge graph model, and input the recommended entity into the cross-compression unit through the knowledge graph model; The probability determination module includes: The account feature extraction unit is used to extract features from the account feature data using a multilayer perceptron through the recommendation model to obtain account features corresponding to the user account. A tag feature extraction unit is used to input the recommendation tag into the cross-compression unit through the recommendation model to obtain the tag features of the recommendation tag, wherein the tag features are determined by the cross-compression unit based on the recommendation tag and the recommendation entity; The click probability determination unit is used to obtain the click probability of the user account for the recommended tag based on the account characteristics and the tag characteristics through the recommendation model.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images